4 #include <rdma/ib_verbs.h>
5 #include <rdma/rdma_cm.h>
6 #include <linux/interrupt.h>
8 #include <linux/slab.h>
10 #include "rdma_transport.h"
12 #define RDS_FMR_SIZE 256
13 #define RDS_FMR_POOL_SIZE 8192
15 #define RDS_IB_MAX_SGE 8
16 #define RDS_IB_RECV_SGE 2
18 #define RDS_IB_DEFAULT_RECV_WR 1024
19 #define RDS_IB_DEFAULT_SEND_WR 256
21 #define RDS_IB_DEFAULT_RETRY_COUNT 2
23 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
25 #define RDS_IB_RECYCLE_BATCH_COUNT 32
27 extern struct rw_semaphore rds_ib_devices_lock;
28 extern struct list_head rds_ib_devices;
31 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
32 * try and minimize the amount of memory tied up both the device and
33 * socket receive queues.
35 struct rds_page_frag {
36 struct list_head f_item;
37 struct list_head f_cache_entry;
38 struct scatterlist f_sg;
41 struct rds_ib_incoming {
42 struct list_head ii_frags;
43 struct list_head ii_cache_entry;
44 struct rds_incoming ii_inc;
47 struct rds_ib_cache_head {
48 struct list_head *first;
52 struct rds_ib_refill_cache {
53 struct rds_ib_cache_head __percpu *percpu;
54 struct list_head *xfer;
55 struct list_head *ready;
58 struct rds_ib_connect_private {
59 /* Add new fields at the end, and don't permute existing fields. */
64 __be16 dp_protocol_minor_mask; /* bitmask */
67 __be32 dp_credit; /* non-zero enables flow ctl */
70 struct rds_ib_send_work {
73 struct ib_send_wr s_wr;
74 struct ib_rdma_wr s_rdma_wr;
75 struct ib_atomic_wr s_atomic_wr;
77 struct ib_sge s_sge[RDS_IB_MAX_SGE];
78 unsigned long s_queued;
81 struct rds_ib_recv_work {
82 struct rds_ib_incoming *r_ibinc;
83 struct rds_page_frag *r_frag;
84 struct ib_recv_wr r_wr;
85 struct ib_sge r_sge[2];
88 struct rds_ib_work_ring {
98 struct rds_ib_connection {
100 struct list_head ib_node;
101 struct rds_ib_device *rds_ibdev;
102 struct rds_connection *conn;
104 /* alphabet soup, IBTA style */
105 struct rdma_cm_id *i_cm_id;
107 struct ib_cq *i_send_cq;
108 struct ib_cq *i_recv_cq;
111 struct rds_ib_work_ring i_send_ring;
112 struct rm_data_op *i_data_op;
113 struct rds_header *i_send_hdrs;
115 struct rds_ib_send_work *i_sends;
116 atomic_t i_signaled_sends;
119 struct tasklet_struct i_recv_tasklet;
120 struct mutex i_recv_mutex;
121 struct rds_ib_work_ring i_recv_ring;
122 struct rds_ib_incoming *i_ibinc;
124 struct rds_header *i_recv_hdrs;
126 struct rds_ib_recv_work *i_recvs;
127 u64 i_ack_recv; /* last ACK received */
128 struct rds_ib_refill_cache i_cache_incs;
129 struct rds_ib_refill_cache i_cache_frags;
132 unsigned long i_ack_flags;
133 #ifdef KERNEL_HAS_ATOMIC64
134 atomic64_t i_ack_next; /* next ACK to send */
136 spinlock_t i_ack_lock; /* protect i_ack_next */
137 u64 i_ack_next; /* next ACK to send */
139 struct rds_header *i_ack;
140 struct ib_send_wr i_ack_wr;
141 struct ib_sge i_ack_sge;
143 unsigned long i_ack_queued;
145 /* Flow control related information
147 * Our algorithm uses a pair variables that we need to access
148 * atomically - one for the send credits, and one posted
149 * recv credits we need to transfer to remote.
150 * Rather than protect them using a slow spinlock, we put both into
151 * a single atomic_t and update it using cmpxchg
155 /* Protocol version specific information */
156 unsigned int i_flowctl:1; /* enable/disable flow ctl */
158 /* Batched completions */
159 unsigned int i_unsignaled_wrs;
162 /* This assumes that atomic_t is at least 32 bits */
163 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
164 #define IB_GET_POST_CREDITS(v) ((v) >> 16)
165 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
166 #define IB_SET_POST_CREDITS(v) ((v) << 16)
168 struct rds_ib_ipaddr {
169 struct list_head list;
173 struct rds_ib_device {
174 struct list_head list;
175 struct list_head ipaddr_list;
176 struct list_head conn_list;
177 struct ib_device *dev;
179 struct rds_ib_mr_pool *mr_pool;
180 unsigned int fmr_max_remaps;
181 unsigned int max_fmrs;
183 unsigned int max_wrs;
184 unsigned int max_initiator_depth;
185 unsigned int max_responder_resources;
186 spinlock_t spinlock; /* protect the above */
188 struct work_struct free_work;
191 #define ibdev_to_node(ibdev) dev_to_node(ibdev->dma_device)
192 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
194 /* bits for i_ack_flags */
195 #define IB_ACK_IN_FLIGHT 0
196 #define IB_ACK_REQUESTED 1
198 /* Magic WR_ID for ACKs */
199 #define RDS_IB_ACK_WR_ID (~(u64) 0)
201 struct rds_ib_statistics {
202 uint64_t s_ib_connect_raced;
203 uint64_t s_ib_listen_closed_stale;
204 uint64_t s_ib_tx_cq_call;
205 uint64_t s_ib_tx_cq_event;
206 uint64_t s_ib_tx_ring_full;
207 uint64_t s_ib_tx_throttle;
208 uint64_t s_ib_tx_sg_mapping_failure;
209 uint64_t s_ib_tx_stalled;
210 uint64_t s_ib_tx_credit_updates;
211 uint64_t s_ib_rx_cq_call;
212 uint64_t s_ib_rx_cq_event;
213 uint64_t s_ib_rx_ring_empty;
214 uint64_t s_ib_rx_refill_from_cq;
215 uint64_t s_ib_rx_refill_from_thread;
216 uint64_t s_ib_rx_alloc_limit;
217 uint64_t s_ib_rx_credit_updates;
218 uint64_t s_ib_ack_sent;
219 uint64_t s_ib_ack_send_failure;
220 uint64_t s_ib_ack_send_delayed;
221 uint64_t s_ib_ack_send_piggybacked;
222 uint64_t s_ib_ack_received;
223 uint64_t s_ib_rdma_mr_alloc;
224 uint64_t s_ib_rdma_mr_free;
225 uint64_t s_ib_rdma_mr_used;
226 uint64_t s_ib_rdma_mr_pool_flush;
227 uint64_t s_ib_rdma_mr_pool_wait;
228 uint64_t s_ib_rdma_mr_pool_depleted;
229 uint64_t s_ib_atomic_cswp;
230 uint64_t s_ib_atomic_fadd;
233 extern struct workqueue_struct *rds_ib_wq;
236 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
239 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
240 struct scatterlist *sglist,
241 unsigned int sg_dma_len,
244 struct scatterlist *sg;
247 for_each_sg(sglist, sg, sg_dma_len, i) {
248 ib_dma_sync_single_for_cpu(dev,
249 ib_sg_dma_address(dev, sg),
250 ib_sg_dma_len(dev, sg),
254 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
256 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
257 struct scatterlist *sglist,
258 unsigned int sg_dma_len,
261 struct scatterlist *sg;
264 for_each_sg(sglist, sg, sg_dma_len, i) {
265 ib_dma_sync_single_for_device(dev,
266 ib_sg_dma_address(dev, sg),
267 ib_sg_dma_len(dev, sg),
271 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
275 extern struct rds_transport rds_ib_transport;
276 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
277 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
278 extern struct ib_client rds_ib_client;
280 extern unsigned int fmr_message_size;
281 extern unsigned int rds_ib_retry_count;
283 extern spinlock_t ib_nodev_conns_lock;
284 extern struct list_head ib_nodev_conns;
287 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
288 void rds_ib_conn_free(void *arg);
289 int rds_ib_conn_connect(struct rds_connection *conn);
290 void rds_ib_conn_shutdown(struct rds_connection *conn);
291 void rds_ib_state_change(struct sock *sk);
292 int rds_ib_listen_init(void);
293 void rds_ib_listen_stop(void);
294 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
295 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
296 struct rdma_cm_event *event);
297 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
298 void rds_ib_cm_connect_complete(struct rds_connection *conn,
299 struct rdma_cm_event *event);
302 #define rds_ib_conn_error(conn, fmt...) \
303 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
306 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
307 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
308 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
309 void rds_ib_destroy_nodev_conns(void);
310 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
311 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
312 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
313 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
314 struct rds_sock *rs, u32 *key_ret);
315 void rds_ib_sync_mr(void *trans_private, int dir);
316 void rds_ib_free_mr(void *trans_private, int invalidate);
317 void rds_ib_flush_mrs(void);
318 int rds_ib_fmr_init(void);
319 void rds_ib_fmr_exit(void);
322 int rds_ib_recv_init(void);
323 void rds_ib_recv_exit(void);
324 int rds_ib_recv(struct rds_connection *conn);
325 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic);
326 void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
327 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp);
328 void rds_ib_inc_free(struct rds_incoming *inc);
329 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
330 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
331 void rds_ib_recv_tasklet_fn(unsigned long data);
332 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
333 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
334 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
335 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
336 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
337 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
340 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
341 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
342 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
343 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
344 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
345 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
346 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
347 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
348 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
349 extern wait_queue_head_t rds_ib_ring_empty_wait;
352 void rds_ib_xmit_complete(struct rds_connection *conn);
353 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
354 unsigned int hdr_off, unsigned int sg, unsigned int off);
355 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
356 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
357 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
358 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
359 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
360 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
361 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
362 u32 *adv_credits, int need_posted, int max_posted);
363 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
366 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
367 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
368 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
372 int rds_ib_sysctl_init(void);
373 void rds_ib_sysctl_exit(void);
374 extern unsigned long rds_ib_sysctl_max_send_wr;
375 extern unsigned long rds_ib_sysctl_max_recv_wr;
376 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
377 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
378 extern unsigned long rds_ib_sysctl_max_recv_allocation;
379 extern unsigned int rds_ib_sysctl_flow_control;